Speaker
Description
Magnetic resonance imaging (MRI) is central to diagnosis, longitudinal monitoring, and response assessment in brain tumors, motivating predictive models that operate not only on volumetric trajectories but directly in the imaging domain. While mechanistic models capture individualized tumor dynamics and treatment effects, they compress the spatial and anatomical information contained in radiographic data; conversely, generative AI solutions such as guided denoising diffusion models, enable anatomically coherent prediction of future MR images and tumor growth probability maps, but require temporal structure that is often difficult to learn from limited follow-up data. Hybrid mechanistic-learning approaches are therefore particularly suitable in this application, where longitudinal imaging is sparse and irregularly sampled, especially in rare pediatric tumors. By coupling biologically plausible ordinary differential equation models of tumor growth with guided denoising diffusion models, these frameworks combine interpretable estimates of future tumor burden with patient-specific image synthesis and spatially resolved progression maps. This integration supports biologically informed, anatomically meaningful prediction of spatio-temporal tumor evolution and provides a foundation for improved response assessment, treatment monitoring, and radiotherapy planning.
